Prompt Engineering Author: Lee Boonstra Prompt Engineering February 2025 2 Acknowledgements Content contributors Michael Sherman Yuan Cao Erick Armbrust Anant Nawalgaria Antonio Gulli Simone Cammel Grace Mollison Technical Writer Joey Haymaker Designer Michael Lanning Introduction 6 Prompt engineering 7 LLM output configuration 8 Output length 8 Sampling controls 9 Temperature 9 Top-K and (CoT) 29 Self-consistency 32 Tree of Thoughts (ToT) 36 ReAct (reason & act) 37 Automatic Prompt Engineering 40 Code prompting 42 Prompts for writing code 42 Prompts for explaining code 44 Prompts

serve a much larger batch size. We evaluate the generation throughput of DeepSeek-V2 based on the prompt and generation length distribution from the actually deployed DeepSeek 67B service. On a single node per second, which is 5.76 times the maximum generation throughput of DeepSeek 67B. In addition, the prompt input throughput of DeepSeek-V2 exceeds 100K tokens per second. 4. Alignment 4.1. Supervised Fine-Tuning an instruction-following evaluation (IFEval) (Zhou et al., 2023) for DeepSeek-V2 Chat (SFT), using prompt-level loose accuracy as the metric. Moreover, we employ LiveCodeBench (Jain et al., 2024) questions

new category of LLM-powered systems known as agents. This guide is designed for product and engineering teams exploring how to build their first agents, distilling insights from numerous customer deployments customer service for example, routines can roughly map to individual articles in your knowledge base. Prompt agents to break 
 down tasks Providing smaller, clearer steps from dense resources 
 helps minimize like o1 or o3-mini, to automatically generate instructions from existing documents. Here’s a sample prompt illustrating this approach: Unset 1 “You are an expert in writing instructions for an LLM agent.

Stanford University… 1: AI ‘Winter’ was a term used by Nils J. Nilsson, the Kumagai Professor of Engineering in computer science at Stanford University, to describe the period during which AI continued to would understand goals, generate plans, and self-correct in real time. They could drive research, engineering, education, and logistics workflows with little to no human oversight – handling ambiguity and databases. Model development = frameworks for modeling & training, inference optimization, dataset engineering, & model evaluation. Application development = custom AI-powered applications (varied use cases)

献，但不是 主要的考虑因素，只有在产品性能差别不大时，才会选择对开源社区贡献大的供应商。  When companies buy open source products, the engineering team leader (technical director/architect / TL) selects most of them. In the same type of software hope the follow-up can be achieved in the learning of computers, compiling principles, software engineering, and other theoretical knowledge at the same time so that students learn to master the open source companies buy open source products (commercial products based on open source projects), most engineering team leaders (technical directors/architects/TLs) will select the product. Half of them will

Global AI Adoption Product Note: With deep research, ChatGPT can do work independently. Give it a prompt, and it can synthesize hundreds of online sources to create comprehensive, PhD-level reports. This Developer resources are the main bottleneck and growth inhibitor in many organizations. 
 When engineering teams are overwhelmed, it slows innovation and creates an insurmountable backlog of apps and ideas

的终极价值是个性化。

海外独⻆兽：OpenAI
其实也有⼀定的
long-context
了。
杨植麟：它还没有把⽤⼾的交互过程真正视为个性化的场景。⽐如，如果我们去
ChatGPT
prompt
某 个东西，
不管是今天还是明天，只要模型版本相同，可能效果基本上差不多，这就是我说的缺乏个性 化。 最终所有东西都是指令遵循。只不过你的指令会越来越复杂。今天你的指令⼀开始可能是
10
个词，但 致效果的追求，所 以市场⾥还存在
Midjourney
之类公司的机会。但是
AGI
的通⽤性⾜够强⼤时，很多⽤⼾也会转移
⸺
如果今天我把
Photoshop
整个软件都重新封装成⼀个
prompt，它变成⼤家⼀个外包的全能设计师， 那会有更少的⼈⽤
Midjourney。
Midjourney
今天的地位在于它通过先发优势让⻜轮跑起来了。⽐较
tricky
的是未来还会不会有这种时 我会从两个例⼦来展开介绍，到底什么是真实的规模化。
我们认为，Transformer是新时代的计算 机。这跟⽼的计算机不⼀样，⽼的计算机可能是你通过编码⽅式实现⼀个确定性的需求，但在 Transformer上，你可能是通过Prompt作为编程语⾔，⽤数据作为桥梁去嫁接你的产品和研发。
在这种新的计算范式下⾯，它会产⽣新的计算，会产⽣新的内存。⽐如说参数数量可能就会决定计算 复杂度，上下⽂的⻓度就会决定内存⼤⼩。有了很⼤

printing in CMYK color space): Scribus [https://scribus.net] CAD (parametrical construction, engineering): LibreCAD (2D) [https://librecad.org/] Painting (raster graphics): Krita [https://krita.org] changes. Adding a Live Path Effect to a Path When you don’t have anything selected, the dialog will prompt you to choose a suitable object (a path, a shape, or a group of paths) before you can proceed. With

printing in CMYK color space): Scribus [https://scribus.net] CAD (parametrical construction, engineering): LibreCAD (2D) [https://librecad.org/] Painting (raster graphics): Krita [https://krita.org] Effects from the bottom of the Path menu entry. When you don’t have a path selected, the dialog will prompt you to choose one before you can proceed. Click on the plus sign to see a list of available path

printing in CMYK color space): Scribus [https://scribus.net] CAD (parametrical construction, engineering): LibreCAD (2D) [https://librecad.org/] Painting (raster graphics): Krita [https://krita.org] changes. Adding a Live Path Effect to a Path When you don’t have anything selected, the dialog will prompt you to choose a suitable object (a path, a shape, or a group of paths) before you can proceed. With